arXiv:1501.07060v1 [math.PR] 28 Jan 2015

Transcription

arXiv:1501.07060v1 [math.PR] 28 Jan 2015
The first-passage time of the Brownian motion
to a curved boundary: an algorithmic approach∗
Samuel Herrmann1 and Etienne Tanré2
arXiv:1501.07060v1 [math.PR] 28 Jan 2015
January 29, 2015
Abstract
Under some weak conditions, the first-passage time of the Brownian
motion to a continuous curved boundary is an almost surely finite stopping
time. Its probability density function (pdf) is explicitly known only in few
particular cases. Several mathematical studies proposed to approximate
the pdf in a quite general framework or even to simulate this hitting time
using a discrete time approximation of the Brownian motion. The authors
study a new algorithm which permits to simulate the first-passage time
using an iterating procedure. The convergence rate presented in this paper
suggests that the method is very efficient.
Key words and phrases: first-passage time, Brownian motion, potential theory, randomized algorithm.
2010 AMS subject classifications: primary 65C05; secondary 65N75, 60G40.
Introduction
Modeling biological or physical systems often requires to handle with onedimensional diffusion processes. The marginal probability distribution of such
processes, at a fixed time, permits to describe quite precisely the model. Nevertheless, in many applications, this information is insufficient and the description
of the whole paths becomes crucial. This is namely the case for variety problems
related to neuronal sciences, financial derivatives with barriers, ruin probability
of an insurance fund, optimal stopping problems,... In these frameworks, the
main task is the description of the first passage time densities for time-dependent
boundaries (for level-crossing problems see [1, 2]).
For instance, let us focus our attention on a simple interpretation of the
neural transmission. When a neuron is stimulated by pressure, heat, light, or
chemical information, its membrane voltage changes as time elapses and, as soon
as it reaches a constant threshold, the depolarization phenomenon occurs and
1 Institut de Mathématiques de Bourgogne, UMR CNRS 5584, Université de Bourgogne,
B.P. 47 870 21078 Dijon Cedex, France [email protected]
2 Inria, Equipe Projet TOSCA, 2004 route des Lucioles, BP93, 06902 Sophia-Antipolis,
France, [email protected]
∗ This work has been supported by the Agence National de la Recherche through the ANR
Project MANDy “Mathematical Analysis of Neuronal Dynamics”, ANR-09-BLAN-0008-01.
1
the voltage is reset to a resting potential. The family of integrate-and-fire spiking
neuron models is based on this simple interpretation. The firing time therefore
corresponds to the first-passage time of the membrane potential, represented by
a stochastic mean-reverting process (usually the Ornstein-Uhlenbeck process)
to the neural threshold (for introduction of noise in neuron systems see Part I
Chapter 5 in [9], for the integrate-and-fire model see Chapter 10 in [8]).
Our main motivation is to emphasize an algorithmic approach in order to
approximate the first-passage time of the Brownian motion to curved boundaries. Using simple time transformations will also permit to apply the results
to the Ornstein-Uhlenbeck process (see Section 3).
In order to describe approximations of the first-passage time of the Brownian
motion, we assume that this stopping time is almost surely finite. In this way, we
introduce particular conditions for this property to be satisfied. Let us consider
a continuous function ϕ : R+ → R satisfying the following hypothesis:
ϕ(0) > 0 and
lim sup √
t→∞
ϕ(t)
< 1.
2t log log t
(H1)
We then define the hitting time
τϕ = inf{t > 0 : Bt = ϕ(t)}
(0.1)
where (Bt , t ≥ 0) stands for a standard one-dimensional Brownian motion.
Under (H1), the a.s. finiteness of τϕ is an obvious consequence of the law of the
iterated logarithm (see e.g.[11, Th.9.23 p.112]). It is quite difficult to obtain
precise informations about this stopping time in general situations.
Durbin [6, 7] proposed to approximate the first-passage distribution p(t) dt =
P(τϕ ∈ dt) of the Brownian motion as follows: p can be represented by a serie
expansion
k
X
p(t) =
(−1)j−1 qj (t) + (−1)k rk (t), k ≥ 1,
j=1
where qj for 1 ≤ j ≤ k and rk are defined by multiple integrals depending on the
boundary ϕ. The approximation simply consists in truncating the series. Let
us note that the first term corresponds in fact to the tangent approximation of
Strassen [15] and Daniels [4]. The convergence of the series and the error bounds
can be precised if the curved boundary is wholly concave or wholly convex.
For particular cases, the probability density function can be computed explicitly. Lerche [12] used the method of images in order to obtain explicit expressions of the p.d.f. p. However, in practice, for general boundaries, the
expression emphasized by Durbin does not permit to simulate easily the hitting
time.
One way to approximate τϕ and especially to compute the probability for the
hitting time to be smaller than some given T > 0, is to use a time discretization
of the Brownian motion on [0, T ]. The time interval is then split into n small
intervals of the kind [(k − 1)T /n, kT /n], with 1 ≤ k ≤ n. At each endpoint
kT /n, the event BkT /n < ϕ(kT /n) has to be tested and at the same time we
need to know if, given the Brownian values at the endpoints, the Brownian paths
on the small intervals (the conditional distribution corresponds therefore to the
Brownian bridge one) hit the curved boundary. The probability not to hit the
boundary on a small interval for the Brownian bridge can be approximated [10].
2
This method can become onerous if the observed time interval [0, T ] is large.
Let us also note that the time-splitting can be replaced by a space-splitting
namely for the first time the Brownian motion exits from a given interval [3].
The aim of this study is to present a new method of approximation of τϕ
by some families (τϕ )≥0 which satisfy exact simulation and such that τϕ converges toward τϕ in distribution as tends to 0. Stopping such sequence at a
small level = 0 induces the error term in the approximation. Two different
families of sequences will be emphasized and the associated convergence rates
are estimated. The first algorithm developed in Section 1 concerns increasing
curved boundaries and the second one, Section 2, permits to deal with quite
general boundaries provided that its derivative is bounded. In the last section,
we present different examples in order to illustrate the algorithm efficiency.
1
First-passage time to non-decreasing boundaries
Let us assume that the boundary ϕ satisfies (H1) and that the following additional conditions hold
ϕ : R+ → R is a non-decreasing C 1 -continuous function,
√
2ϕ0 (t) 1 + t ≤ 1, ∀t ≥ 0.
(H2)
(H3)
We introduce the algorithm associated to the hitting time τϕ defined by (0.1).
Algorithm (A1). Let > 0 be a small parameter and (Gn )n≥0 a sequence of
independent standard Gaussian distributed random variables.
Initialization: T0 = 0, T1 = 0, T2 = (ϕ(0)/G0 )2 and N = 1.
While ϕ(T2 ) − ϕ(T1 ) > do:
(
(T0 , T1 , T2 ) ← T1 , T2 , T2 + (ϕ(T2 ) − ϕ(T1 ))2 /G2N
(1.1)
N ← N + 1.
Outcome: τϕ ← T2 and N .
Let us just note that Algorithm (A1) is very simple to use since each step
only requires one Gaussian distributed random variable. Moreover it is a approximation of the first-passage time:
Theorem 1.1.
1. Let us assume that the boundary function ϕ satisfies (H1),
(H2) and (H3) then the random variable τϕ defined in Algorithm (A1)
converges in distribution towards τϕ defined by (0.1) as tends to zero.
More precisely
√
3 F (t − ) − √
≤ F (t) ≤ F (t), for any t ≥ ,
(1.2)
2π
where F (resp. F ) is the cumulative distribution function of τϕ (resp.
τϕ ).
2. There exists a constant C > 0 such that the random number of iterations
N defined in Algorithm (A1) satisfies:
p
E[Nε ] ≤ C | log |.
(1.3)
3
The parameter describes the precision of the approximation. The number
of steps in the Algorithm (A1) is very small (even smaller than usual results
obtained for algorithms based on random walks on spheres, which are close
to Algorithm (A1), see [14]) : in fact the constant appearing in (1.3) can be
explicitly computed: for any constant 0 < κ < 1/2, there exists 0 (κ) > 0 such
that (1.3) is satisfied as soon as < 0 , with the particular constant
√
2 2
1
√
, m = log(4) +
µ
C=
mκ
π
and
Z
µ=
∞
2
(log |x|) e−x dx.
(1.4)
0
The proof of Theorem 1.1 is based on a main argument developed in the following
proposition: each step of Algorithm (A1) has to be related to a particular part
of the Brownian paths before hitting the boundary.
Proposition 1.2. Let (Bt , t ≥ 0) be a standard one-dimensional Brownian
motion. We define the following sequence of stopping times: s0 = T0 = 0 and
for any n ≥ 1:
n
o
sn := inf t ≥ 0 : Bt+Tn−1 = ϕ(Tn−1 )
and Tn := s1 + . . . + sn ,
(1.5)
where the function ϕ satisfies (H1), (H2) and (H3). Then the following properties hold:
1. (Tn )n≥0 is a non-decreasing sequence which almost surely converges towards τϕ .
2. Let n ≥ 1, then the probability distribution of sn+1 given the σ-algebra
Fn := σ{s1 . . . , sn } is identical as (ϕ(Tn ) − ϕ(Tn−1 ))2 /G2n where (Gn )n≥0
is a sequence of independent standard Gaussian random variables. More(d)
over s1 = (ϕ(0)/G0 )2 .
3. Let M := inf{n ≥ 1 : ϕ(Tn ) − BTn ≤ }, then TM and τϕ , defined in
Algorithm (A1), are identically distributed, so are M and N .
Let us note that the mean of each random variable sn defined by (1.5) is infinite since E[G−2 ] = +∞ where G is a standard Gaussian variable. Proposition
1.2 suggests that the first-passage time can be obtained as a sum of positive
random variables of infinite average, we easily deduce E[τϕ ] = +∞. In the
particular case of increasing boundaries ϕ, the sum has infinitely many terms.
Proof of Proposition 1.2.
Step 1. By construction, the sequence (Tn )n≥0 is non-decreasing and nonnegative: it converges almost surely to T∞ . Since ϕ is a non-decreasing boundary, Tn ≤ τϕ for any n ≥ 0. In particular T∞ is less than τϕ which is a finite
stopping time due to the law of the iterated logarithm, see (H1) followed by
discussion. Consequently, the random variable BT∞ is well defined. Since ϕ is
non-decreasing, we get BTn = ϕ(Tn−1 ) for any n ≥ 1. Taking the large n limit
leads to BT∞ = ϕ(T∞ ), the Brownian paths and the function ϕ being continuous. We deduce that T∞ = τϕ .
4
Step 2. Let us first consider the stopping time s1 . Using the reflection principle
of the Brownian paths and a scaling property, we obtain:
P(s1 > t) = P sup Bu < ϕ(0) = P(|Bt | < ϕ(0))
0≤u≤t
= P(B12 < ϕ(0)2 /t) = P(ϕ(0)2 /G20 > t),
t ≥ 0.
The general n-th case can be proven using similar arguments combined with the
Markov property of the Brownian motion:
P(sn+1 > t|Fn ) = P
sup
Bu < ϕ(Tn )Fn
Tn ≤u≤Tn +t
= P sup Bu+Tn − BTn < ϕ(Tn ) − ϕ(Tn−1 )Fn
0≤u≤t
˜u < ϕ(Tn ) − ϕ(Tn−1 )Fn ,
= P sup B
0≤u≤t
˜ is a Brownian motion independent of Fn .
where B
Step 3. Using the results developed in Step 2, we observe that (sn )n∧M and
the sequence of values T2 , defined in Algorithm (A1), have the same distribution.
It is therefore obvious that TM and τϕ are identically distributed. Indeed the
stopping time can be rewritten as follows:
M = inf{n ≥ 1 : ϕ(Tn ) − ϕ(Tn−1 ) ≤ }.
(1.6)
Proof of Theorem 1.1.
Step 1. Let us recall that Tn is defined by (1.5). By Proposition 1.2, Tn ≤ τϕ
for any n ≥ 0 and in particular TM ≤ τϕ . Hence
P(TM ≤ t) ≥ P(τϕ ≤ t),
∀t ≥ 0.
Since τϕ has the same distribution as TM , we obtain
F (t) ≥ F (t),
∀t ≥ 0,
(1.7)
where F and F are the associated cumulative distribution functions. Let us
now prove the second bound in (1.2). For t ≥ ,
F (t − ) = P(τϕ ≤ t − ) = P(TM ≤ t − )
≤ P(TM ≤ t − , τϕ > t) + P(τϕ ≤ t)
≤ P(|TM − τϕ | > ) + F (t).
(1.8)
Combining the Markov property of the Brownian motion and the reflection
5
principle leads to
P := P(|TM − τϕ | > ) ≤ 1 − P sup BTM +u ≥ sup ϕ(TM + u)
0≤u≤
0≤u≤
≤ 1 − P sup BTM +u − BTM ≥ sup ϕ(TM + u) − ϕ(TM ) + 0≤u≤
0≤u≤
≤ 1 − P sup BTM +u − BTM ≥ ϕ(TM + ) − ϕ(TM ) + 0≤u≤
˜u ≥
≤ 1 − P sup B
0≤u≤
˜ | ≥
≤ 1 − P |B
ϕ0 (θ) + sup
TM ≤θ≤TM +
ϕ0 (θ) + .
sup
TM ≤θ≤TM +
Using Hypothesis (H3) and straightforward computations permits to obtain
r
˜ | ≥ 3/2) ≤ 3 .
P(|TM − τϕ | > ) ≤ 1 − P(|B
(1.9)
2π
The lower bound in (1.2) holds due to both (1.8) and (1.9).
Step 2. Let us now focus our attention to the efficiency of this algorithm. We
need to estimate the number of steps which depends on the small parameter .
Using the third result presented in Proposition 1.2 on one hand and (1.6) on
the other hand, we obtain
P(N > n) = P(M > n) = P(ϕ(T1 ) − ϕ(T0 ) > , . . . , ϕ(Tn ) − ϕ(Tn−1 ) > ).
Hypothesis (H3) implies
P(N > n) ≤ P(s1 > 2, . . . , sn > 2).
(1.10)
Step 2.1. Let us first estimate the previous upper-bound. We introduce a sequence of independent standard Gaussian random variables (Gn )n≥0 and define
Xn =
log(4G2n ),
Ξn =
n
X
k=0
Xk
and Zn =
n
X
Ξk .
(1.11)
k=0
Let us define Π(n, ) := P(sn > 2). By Proposition 1.2, we know that the
random variables sn+1 are related to Gn and therefore
Π(1, ) = P(2G20 < ϕ(0)2 ) = P log(4G20 ) < − log() + log(2) + 2 log ϕ(0)
= P Z0 < − log() + log(2) + 2 log ϕ(0) .
Let us prove that, for n ≥ 1, we have the general formula:
Π(n, ) ≤ P Zn−1 < − log() + (2n − 1) log(2) + (2n) log ϕ(0) .
(1.12)
By Proposition 1.2, we have for n ≥ 2,
Π(n, ) = P (ϕ(Tn−1 ) − ϕ(Tn−2 ))2 > 2G2n−1 .
(1.13)
6
Since ϕ is a non decreasing function satisfying Hypothesis (H3), the following
upper-bound holds for n ≥ 2:
sn−1
Tn−1 − Tn−2
≤ √
ϕ(Tn−1 ) − ϕ(Tn−2 ) ≤ p
.
2
1
+ sn−2
2 1 + Tn−2
(1.14)
Hence for n = 2, (1.13) and (1.14) imply
s2
Π(2, ) ≤ P 12 > 2G21 = P (2G21 )(2G20 )2 < ϕ(0)4
2
= P(2X0 + X1 < − log() + 3 log(2) + 4 log ϕ(0))
= P(Z1 < − log() + 3 log(2) + 4 log ϕ(0)).
Using the lower-bound 1+sn−1 ≥ sn−1 and similar arguments as those developed
previously, the general case is expressed as follows:
s3
s2n−1
n−2
2
2
4
≤
P
>
2G
>
2G
G
n−1
n−1 n−2
22 (1 + sn−2 )
22 24 s2n−3
sn−1
2(n−2)
≤ P 2n−2 > 222 24 . . . 22(n−2) G2n−1 G4n−2 . . . G2
s1
s2 n−1 1
2(n−2)
1
2 4
2(n−2) 2
4
≤P
>
22
2
.
.
.
2
G
G
.
.
.
G
n−1 n−2
2
22 G21
sn−2
1
≤ P ϕ(0)2n > 222 24 . . . 22(n−1) G2n−1 G4n−2 . . . G2n
0
≤ P Zn−1 < − log() + (2n − 1) log(2) + (2n) log ϕ(0) .
Π(n, ) ≤ P
Step 2.2. By (1.10) and the arguments developed in Step 2.1, we obtain
P(N > n) ≤ P(sn > 2) ≤ P(Zn−1 − EZn−1 < η(, n) − EZn−1 ),
where
η(, n) := − log() + (2n − 1) log(2) + (2n) log ϕ(0).
Let us observe that, for any n ≥ 0, m := E[Xn ] = log(4) +
defined by (1.4). Hence
E[Zn ] =
n
X
k=0
E[Ξn ] =
n X
k
X
E[Xj ] = m
k=0 j=0
n
X
(k + 1) =
k=0
√
2√ 2
µ
π
> 0 where µ is
m(n + 1)(n + 2)
.
2
Thus, for n large enough, η(, n) − EZn−1 < 0. Introducing dn := |mn(n +
1)/2 − η(, n)|, we observe that, for any 0 < κ < 1/2 there exists ℵ(κ, ) ∈ N
such that dn > mn2 (1/2 − κ) for n sufficiently large that is n ≥ ℵ(κ, ). After
straightforward computations, we can choose
r
j | log(2)| 1
log(2ϕ(0)) k
ℵ(κ, ) :=
+
−
(1.15)
+ 1.
mκ
2κ
mκ
Markov’s inequality leads to
P(N > n) ≤ P(|Zn−1 − E[Zn−1 ]| > dn ) ≤
7
E[(Zn−1 − E[Zn−1 ])4 ]
.
d4n
(1.16)
Let us note that X j := Xj − m are i.i.d. random variables with finite moments
k
of any order. We denote mk := E[X j ]. Therefore we obtain
k
4 i
h n−1
4 i
h n−1
X
XX
Xj
=E
Zn−1 := E[(Zn−1 − E[Zn−1 ])4 ] = E
(n − j)X j
j=0
k=0 j=0
=
n−1
X
X
j=0
0≤j<k≤n−1
(n − j)4 m4 + 2
(n − j)2 (n − k)2 m22
m2
m4
n(n + 1)(6n3 + 9n2 + n − 1) + 2 n2 (n + 1)2 (2n + 1)2 .
(1.17)
30
36
Hence, there exist a constant C0 > 0 such that E[(Zn−1 − E[Zn−1 ])4 ] ≤ C0 n6 .
Combining the previous inequality with (1.15) and (1.16) leads to
≤
P(N > n) ≤
C0
4
m (1/2 −
κ)4
1
,
n2
for n ≥ ℵ(κ, ).
Consequently, the following upper-bound holds
X
C0
E[N ] =
P(N > n) ≤ ℵ(κ, ) + 4
m (1/2 − κ)4
n≥0
X
n≥ℵ(κ,)
1
.
n2
In order to conclude, it suffices to note that ℵ(κ, ) → ∞ as → 0, the second
term in the previous inequality therefore becomes
small as → 0: the leading
p
term is finally ℵ(κ, ) which is equivalent to | log(2)|/(mκ) by (1.15).
2
First-passage time to boundaries with bounded
derivative
The algorithm presented in Section 1 is simple to achieve (it only requires independent Gaussian
p random variables) and efficient: the averaged number of steps
is of the order | log | where stands for the small parameter appearing in the
rejection sampling (see Theorem 1.1). In order to apply Algorithm (A1) the
curved boundary, the Brownian motion is going to hit, has to satisfies suitable
conditions: (H1), (H2) and (H3). Asking for the monotonicity of the function ϕ
is quite restrictive, that’s why we present an extension of the algorithm which
is of course less efficient (even if the average number of steps is still very small)
but which permits to deal with more general boundaries. Let us introduce the
following assumption: there exist two constants ρ+ > 0 and ρ− > 0 such that
ϕ : R+ → R is a C 1 -continuous function satisfying
sup ϕ0 (t) ≤ ρ+
t≥0
and
inf ϕ0 (t) ≥ −ρ− .
t≥0
(H4)
For such boundaries, we present an algorithm which permits for any K ∈ R+
to approximate the hitting time τϕK = τϕ ∧ K, where τϕ is defined in (0.1). Let
us introduce some notations: the inverse Gaussian distribution of parameters
µ > 0 and λ > 0 will be denoted by I(µ, λ) and is defined by its the probability
distribution function:
r
n λ(x − µ)2 o
λ
f (x) =
exp
−
1{x≥0} .
2πx3
2µ2 x
8
Algorithm (A2). Let > 0 be a small parameter and r > ρ− where ρ− is
defined in (H4).
Initialization: (T, H) = (0, ϕ(0)) and N,K = 0.
ˆ an inverse Gaussian random variable with
While H > and T < K, simulate G
2
distribution I(H/r, H ) and do:

ˆ − ϕ(T ) + r G,
ˆ
 H ← ϕ(T + G)
ˆ
(2.1)
T ← G + T,

N,K ← N,K + 1.
Outcome: τϕ,K ← T ∧ K and N,K .
Algorithm (A2) is quite simple, it only requires the simulation of inverse
Gaussian distributed random variables. Let us recall the following scaling
2
ˆ
ˆ ∼ I(H/r, H 2 ) then H G/r
ˆ
is
property: if G
∼ I(1, rH). Moreover (rG−H)
ˆ
G
Chi-squared distributed with one degree of freedom (the square of a standard
Gaussian random variable). In order to simulate an inverse Gaussian random
variable, we suggest to use the algorithm introduced by Michael, Schucany and
Haas (see [13] or [5, p. 149]). Let us now state the efficiency of Algorithm (A2).
The inverse Gaussian distribution does not permit to argue in a similar way as
in Section 1. That’s why we are going to use the general potential theory in
order to upper-bound the averaged number of steps. This kind of arguments
was already introduced in convergence results associated to the Random Walk
on Spheres algorithm which permits to approximate the solution of the Dirichlet
problem, see for instance [14].
Theorem 2.1.
1. Let us assume that the boundary function ϕ satisfies (H4)
then the random variable τϕ,K defined in Algorithm (A2) converges in
distribution towards τϕK = τϕ ∧ K where τϕ is defined by (0.1) as tends
to zero. More precisely
r
2
F,K (t − ) − (1 + ρ)
≤ FK (t) ≤ F,K (t), for any t ≥ , (2.2)
π
where FK (resp. F,K ) is the cumulative distribution function of τϕK (resp.
τϕ,K ).
2. There exist positive constants a, b, κ0 , κ1 and 0 such that: for any
ρ+ ≤ κ0 and any (K, r) satisfying (r + κ0 )K ≤ κ1 , the random number of iterations N,K defined in Algorithm (A2) satisfies the following
upper bound
E[Nε,K ] ≤ (a + br)| log |, ∀ ≤ 0 .
(2.3)
3. For non increasing functions ϕ: there exists two positive constants a and
0 such that
E[Nε,K ] ≤ ar2 K| log |, ∀ ≤ 0 .
(2.4)
This theorem is based on the following intermediate statement which is a
modification of Proposition 1.2.
9
Proposition 2.2. Let (Bt , t ≥ 0) be a standard one-dimensional Brownian
motion. We introduce the following stopping times: s0 = T0K = 0 and for any
n ≥ 1:
n
o
K
K
sn := inf t ≥ 0 : Bt+Tn−1
= ϕ(Tn−1
) − rt
and TnK := (s1 + . . . + sn ) ∧ K,
(2.5)
where the boundary ϕ satisfies (H4). Then the following properties hold:
1. (TnK )n≥0 is a non-decreasing sequence which almost surely converges towards τϕK .
2. On the event {s1 +· · ·+sn < K}, the probability distribution of sn+1 given
the σ-algebra Fn := σ{T1K , . . . , TnK } is the inverse Gaussian distribution
I(Hn /r, Hn2 ) with
K
Hn := ϕ(TnK ) − ϕ(Tn−1
) + rsn .
(2.6)
3. Let M := inf{n ≥ 1 : ϕ(TnK )−BTnK ≤ }, MK := inf{n ≥ 1, TnK = K},
K
MK
and τϕ,K , defined in Algorithm (A2), are
= M ∧ M . then TMK
identically distributed, so are MK
and N,K .
Proof of Proposition 2.2. The first and the third part of the proof are left to
the reader. They need similar arguments as those presented in Proposition 1.2.
Here the monotonicity property is just replaced by (H4) which permits easily
to prove that TnK ≤ τϕK .
Let us now focus our attention to the second part of the statement. Due to the
K
) − rsn . Hence,
definition of sn+1 and since {TnK < K}, we get BTnK = ϕ(Tn−1
we have
sn+1 = inf{t ≥ 0 : Bt+TnK − BTnK = ϕ(TnK ) − BTnK − rt}
= inf{t ≥ 0 : Wt = Hn − rt},
where Wt = Bt+TnK − BTnK is a standard Brownian motion independent of Fn
and the Fn adapted r.v. Hn is defined by (2.6). The distribution of sn+1 corresponds to the distribution of the first passage time of the standard Brownian
motion with drift at the constant level Hn . The probability distribution is well
known (see, for instance [11, p. 197]):
n (H − rt)2 o
Hn
n
P(sn+1 ∈ dt|Fn ) = √
exp −
dt,
3
2t
2πt
we can consequently identify the inverse Gaussian distribution I(Hn /r, Hn2 ).
Proof of Theorem 2.1.
Step 1. We can prove the convergence in distribution of τϕ,K towards τϕK using
similar arguments as those presented in the proof of Theorem 1.1. The upperbound in (2.2) is an adaptation of (1.7) which requires that TnK ≤ τϕK and that
τϕ,K and TMK
are identically distributed. These conditions are satisfied, see
Proposition 2.2. For the lower-bound in (2.2), we obtain
F,K (t − ) ≤ P(|TMK
− τϕK | > ) + F,K (t),
10
see (1.8) for the details. Let us note that |τMK
− τϕK | > leads to the condition
K
τMK
< K. Hence M = M . Using the Markov property, the following bound
holds:
K
˜u ≥ + sup ϕ(TM + u) − ϕ(TM ) .
P(|TMK
−
τ
|
>
)
≤
1
−
P
sup
B
ϕ
0≤u≤
0≤u≤
˜t , t ≥ 0) stands for a standard Brownian motion independent of TM .
Here (B
Combining Hypothesis (H4) and the reflection principle of the Brownian motion
leads to
r
2
K
˜
,
P(|TM − τϕ | > ) ≤ 1 − P(|B | ≥ (1 + ρ+ )) ≤ (1 + ρ+ )
π
and consequently to the lower bound (2.2).
Step 2. Let us now focus our attention to the averaged number of steps in
Algorithm (A2), denoted by N,K . A rough description of the method: we aim to
construct a Markov chain and to describe the associated potential. The classical
potential theory then permits to obtain the announced bound. We introduce
the Markov chain Rn := (Tn , Hn ) for n ≥ 0. We recall that Tn = s1 + . . . + sn is
defined by (2.5) and Hn by (2.6). The stopping time MK
defined in Proposition
2.2 can also be interpreted as the first time the Markov chain (Rn , n ≥ 0) goes
out of the domain E := [0, K]×], +∞].
Let us consider the function f (x, y) = log(y), defined on E, and denote by
P the infinitesimal generator associated to the Markov chain (Rn )n≥0 . By
Proposition 2.2 and for any (t, h) ∈ E, we obtain
h
i
ˆ − ϕ(t) + r G)
ˆ ,
P f (t, h) = E log(ϕ(t + G)
ˆ is an inverse Gaussian distributed random variable with the following
where G
density function:
h
(h − rx)2
p(x) = √
exp −
, x ≥ 0.
2x
2πx3
ˆ − ϕ(t) ≤ ρ+ G,
ˆ we get
By (H4), ϕ(t + G)
h
rG
ˆ i
ρ+ P f (t, h) − f (t, h) ≤ log 1 +
+ E log
.
r
h
(2.7)
Let us find now an explicit upper bound of P f − f . Using first the change of
variables u = rx/h and secondly u 7→ 1/u, we get
h
rG
rx h
ˆ i Z ∞
(h − rx)2
√
E log
=
log
exp −
dx
h
h
2x
2πx3
r0
Z
hr ∞ log(u)
hr(1 − u)2
=
exp −
du
3/2
2π 0
2u
u
r
Z
hr ∞ (1 − u) log(u)
hr(1 − u)2
=
du.
(2.8)
exp
−
2π 1
2u
u3/2
h
i
ˆ
It is then obvious that E log rhG < 0. Let us now give a more precise upperh
i
ˆ
bound. We set α = hr, then (2.8) emphasizes that E log rhG
only depends
11
Figure 1: Monte Carlo approximation of the function ψ
on the parameter α, this dependence being continuous. Let us therefore denote
this function ψ(α) (see Figure 1 below representing ψ obtained with the MonteCarlo method sample size: 10 000).
Simple computations lead to
r
Z ∞
h
rG
ˆ i
αu2
α
u log(1 + u)
exp
−
ψ(α) := E log
=−
du
(2.9)
h
2π 0
2(1 + u)
(1 + u)3/2
r
Z ∞
α
u log(1 + u)
αu
du
≤−
exp −
3/2
2π 0
2
(1 + u)
Z ∞
1
w log(1 + w/α)
w
≤ −√
exp − dw
3/2
2
(α + w)
2π 0
Z ∞
w log(1 + w/α)
1
w
≤ −√
exp − dw.
2
2π 1/2 (α + w)3/2
Using the inequality (α + w) ≤ (1 + 2α)w, we get
Z
log(1 + (2α)−1 ) ∞ 1
w
√
√ exp − dw
ψ(α) ≤ −
3/2
2
w
(1 + 2α)
2π 1/2
≤−
log(1 + (2α)−1 )
P(G ≥ 1/2),
(1 + 2α)3/2
where G is a standard gaussian r.v. and so P(G ≥ 1/2) ≈ 0.3085
We deduce from the previous upper-bound that limα→0+ ψ(α) = −∞. Moreover
the right hand side is a non decreasing function with respect to the variable α.
Hence
log(3/2)
ψ(α) ≤ − √
P(G ≥ 1/2) ≈ −0.0241, for α ≤ 1.
(2.10)
3 3
Let us observe what happens for large values of the variable α. The Laplace
method implies that
1
ψ(α) ∼ −
as α → ∞.
2α
12
Let us prove now that there exists a constant c > 0 such that
c
ψ(α) ≤ − , for any α ≥ 1.
α
For α ≥ 1, we get
r
Z ∞
αu2
α
u log(1 + u)
exp −
ψ(α) ≤ −
du
3/2
2π 0
2
(1 + u)
r
Z 1
αu2
α
u log(1 + u)
exp
−
≤−
du.
2π 0 (1 + u)3/2
2
(2.11)
Due to the convexity property of the logarithm function (log(1 + u) ≥ log(2)u)
and the Cauchy-Schwarz inequality, we obtain
log(2) 1
E[G2 ] − E[G2 1{G≥√α} ]
ψ(α) ≤ − 3/2
2
α2
q
√ log(2) 1 p
≤ − 3/2
− E[G4 ] P(G ≥ α)
2 √
α2
√
3 −α log(2)
log(2) 1
−
e
≤ − 5/2 (1 − 3e−1 ), for α ≥ 1.
≤ − 3/2
2
2
α2
α2
We deduce that ψ(α) ≤ −c/α with c ≈ 0.0445 when α ≥ 1. Combining both
inequalities (2.10) and (2.11) leads to the existence of a constant c > 0 such
that
1
ψ(α) ≤ −c
∧1 .
(2.12)
α
By (2.7), the following upper-bound holds: for f (x, y) = log(y),
1
ρ+ P f (t, h) − f (t, h) ≤ log 1 +
∧1
−c
r
hr
1
ρ+
≤
−c
∧ 1 , h ≥ 0, t ≥ 0.
(2.13)
r
hr
Due to the definition of ρ+ , we know that
h ≤ ϕ(0) ∨ (r + ρ+ )t ≤ ϕ(0) ∨ (r + ρ+ )K,
where ϕ is the boundary the process has to hit. In other words, there exist two
constants κ0 > 0 and κ1 > 0 such that for any ρ+ ≤κ0 and any (K, r) satisfying
(r + κ0 )K ≤ κ1 the following bound holds ρ+ ≤
P f (t, h) − f (t, h) ≤ −
c
2
1
h
∧ r . Hence:
c
1
∧ r =: −R−1 (r).
2r ϕ(0) ∧ κ1
We deduce that the function g(t, h) defined by g(t, h) = R(r) (f (t, h) − log )
satisfies g(t, h) ≥ 0 for any (t, h) ∈ E and P g(t, h) − g(t, h) ≤ −1 on E. The
potential theory therefore implies:
E[N,K ] ≤ g(0, ϕ(0)) ≤ R(r)(log(ϕ(0)) − log()).
We finally deduce the existence of a > 0 and b > 0 such that E[Nε,K ] ≤
(a + br)| log | for small enough.
For the particular case of a non increasing boundary function it suffices to vanish
ρ+ in (2.13) and to apply the same arguments of the potential theory in order
to get (2.4)
13
3
Examples and numerics.
In this section, we present three different examples which nicely illustrate the
efficiency of these new algorithms (A1) and (A2).
√
Brownian hitting time of ϕ(t) =
3.1
1 + αt
6
6
5
5
mean number of steps
mean number of steps
Let
√ us first consider an application of Theorem 1.1. We observe that ϕ(t) =
1 + αt is an increasing function satisfying (H1), (H2) and (H3) for α ∈ [0, 1].
Consequently Algorithm (A1) converges and permits to obtain an approximation
of the hitting time τϕ . In the figures, we present the link between the averaged
number of steps and which characterizes the approximation error size.
The first figure (resp. the second one) concerns: α = 1 (resp. α = 0.01),
= 0.5n (n is represented on the horizontal axis) and the number of simulation
in order to estimate the averaged number of steps is 10 000.
4
3
2
1
0
4
3
2
1
1
2
3
4
5
6
7
8
9
0
10
1
2
3
4
5
n
6
7
8
9
10
n
(a) α = 1
(b) α = 0.01
n
Figure 2: E(N ): mean
√ number of steps for = 0.5 as a function of n. The
boundary is ϕ(t) = 1 + αt.
Let us now present the approximate distribution of the hitting time.
0.4
0.4
0.3
0.3
pdf
0.5
pdf
0.5
0.2
0.2
0.1
0.1
0
0
2
4
6
8
10
12
14
16
18
0
20
t
0
2
4
6
8
10
12
14
16
18
20
t
(a) α = 1
(b) α = 0.01
Figure 3: Empirical
√ distribution of the approximate first hitting time of the
boundary ϕ(t) = 1 + αt.
14
3.2
Brownian hitting time of ϕ(t) = α + β cos(ωt)
Let us now consider the first time the Brownian motion hits the periodic boundary ϕ(t) = α + β cos(ωt). Since the boundary is not an increasing function, we
shall use Algorithm (A2). Theorem 2.1 ensures that the algorithm converges.
Let us therefore use the Monte-Carlo method in order to estimate precisely the
average number of steps. As explained in the previous section, the simulation
procedure permits to approximate the stopping time τϕ ∧K for some given fixed
time K. Figure 4 illustrates the approximation τϕ by τϕ,K , where the parameters are fixed at α = 3.5, β = 3 and ω = π/2. The maximal time are K = 20 on
one hand and K = 100 on the other hand and the error rate is given by = 0.5n ,
for 1 ≤ n ≤ 10. A sample of 10E8 paths has been simulated to approximate the
mean. We know that the mean number of steps is a decreasing function of and
(a) E(N1/2n ,K ) versus n
1/2n ,K
(b) Distribution of τϕ
. n = 10, K = 20.
Figure 4: Approximation of τϕ with ϕ(t) = 3.5 + 3 cos(πt/2)
an increasing function of K. Figure 5 gives the evolution of the mean number
of steps as a function of the truncation K. In practice, we obtained easily an
Figure 5: E(Nϕ,K ) as a function of K.
impressively accurate approximation of τϕ .
15
3.3
The first time the Ornstein Uhlenbeck process hits the
boundary ϕ(t) = α + β cos(ωt)
The last example concerns the one-dimensional Ornstein-Uhlenbeck process defined as the unique solution of the following stochastic differential equation:
dXt = dBt − λXt dt,
X 0 = x0 ,
(3.1)
where (Bt , t ≥ 0) is the standard Brownian motion. The aim is to approximate
the first passage time through the curved boundary ϕ(t) = α + β cos(ωt) where
ϕ(0) > x0 . Since the Ornstein-Uhlenbeck process can be represented as a timechanged Brownian motion, the question is directly related to the main results
of this study. Indeed the solution of (3.1) is given by
Z t
−λt
λs
Xt = e
x0 +
e dBs , t ≥ 0.
0
Using Levy’s theorem, (Xt , t ≥ 0) has the same distribution as (Yt , t ≥ 0)
defined by
Yt := e−λt x0 + Wu(t) , t ≥ 0,
with u(t) :=
1
2λ
(e2λt − 1) and W a standard Brownian motion. We deduce that
Tϕ := inf{t ≥ 0 : Xt = ϕ(t)}
has the same distribution as
n
o
Tˆϕ := inf t ≥ 0 : e−λt x0 + Wu(t) = ϕ(t)
n
o
−1
= inf u−1 (s) ≥ 0 : Ws = ϕ(u−1 (s))eλu (s) − x0
= u−1 (τψ ),
where
τψ := inf{t ≥ 0 : Wt = ψ(t)},
ψ(t) :=
√
1 + 2λt ϕ
log(1 + 2λt) 2λ
− x0 .
Consequently, in order to simulate the Ornstein-Uhlenbeck hitting time Tϕ ∧ K
for some K, we simply use Algorithm (A2) and propose an approximation of
˜ with K
˜ := u(K) = (e2λK − 1)/(2λ).
the Brownian hitting time τψ ∧ K
Let us note that a straightforward computation leads to the following upperbound:
λα + λβ + ωβ
≤ λα + λβ + ωβ, t ≥ 0.
|ψ 0 (t)| ≤ √
1 + 2λt
In other words, the continuous curve ψ satisfies Hypothesis (H4): Algorithm
(A2) therefore converges and Theorem 2.1 can be applied.
In the following numerical experiences, we will choose r = 0.5 + λα + λβ + ωβ.
Figures 6 and 7 concern the following choice of parameters: x0 = 0, α = 2,
β = 1, ω = π/5, λ = 0.5. We have chosen K = 5 for Figure 6 and K = 10 for
Figure 7. In both cases, the first figure represents the average number of steps
as a function of n where the approximation parameter is chosen as 0.5n , for
n = 1, · · · , 10. The average has been estimated using 5.10E6 simulations. The
second figure represents the distribution of Tϕ ∧ K for n = 10.
16
˜ = (e2λK − 1)/(2λ) increases very fast
We observe that the change of time K
with K and the number becomes quite large when K increases. Note however
that the number of random variables we have to simulate keeps relatively small
in comparaison with the use of a classical stopped Euler scheme usually used to
approximate Tϕ .
(b) Distribution of TϕK, ( = 1/210 ).
(a) E(N1/2n ,K ) versus n
Figure 6: First hitting time of ϕ(t) = α + β cos(ωt) by an Ornstein Uhlenbeck
process solution of (3.1) (α = 2, β = 1, ω = π/5, λ = 0.5, K = 5.)
(b) Distribution of TϕK, ( = 1/210 ).
(a) E(N1/2n ,K ) versus n
Figure 7: First hitting time of ϕ(t) = α + β cos(ωt) by an Ornstein Uhlenbeck
process solution of (3.1) (α = 2, β = 1, ω = π/5, λ = 0.5, K = 10).
References
[1] J. Abrahams. A survey of recent progress on level-crossing problems for
random processes. In IanF. Blake and H.Vincent Poor, editors, Communications and Networks, pages 6–25. Springer New York, 1986.
[2] I. F. Blake and W. C. Lindsey. Level-crossing problems for random processes. IEEE Trans. Information Theory, IT-19:295–315, 1973.
17
[3] Z. A. Burq and O. D. Jones. Simulation of Brownian motion at first-passage
times. Math. Comput. Simulation, 77(1):64–71, 2008.
[4] H. E. Daniels. The maximum size of a closed epidemic. Advances in Appl.
Probability, 6:607–621, 1974.
[5] L. Devroye. Nonuniform random variate generation. Springer-Verlag, New
York, 1986.
[6] J. Durbin. The first-passage density of a continuous Gaussian process to a
general boundary. J. Appl. Probab., 22(1):99–122, 1985.
[7] J. Durbin. The first-passage density of the Brownian motion process to a
curved boundary. J. Appl. Probab., 29(2):291–304, 1992. With an appendix
by D. Williams.
[8] G. B. Ermentrout and D. H. Terman. Mathematical foundations of neuroscience, volume 35 of Interdisciplinary Applied Mathematics. Springer,
New York, 2010.
[9] W. Gerstner and W. M. Kistler. Spiking neuron models. Cambridge University Press, Cambridge, 2002. Single neurons, populations, plasticity.
[10] E. Gobet. Euler schemes and half-space approximation for the simulation
of diffusion in a domain. ESAIM Probab. Statist., 5:261–297 (electronic),
2001.
[11] I. Karatzas and S. E. Shreve. Brownian motion and stochastic calculus,
volume 113 of Graduate Texts in Mathematics. Springer-Verlag, New York,
second edition, 1991.
[12] H. R. Lerche. Boundary crossing of Brownian motion, volume 40 of Lecture
Notes in Statistics. Springer-Verlag, Berlin, 1986. Its relation to the law of
the iterated logarithm and to sequential analysis.
[13] J. R. Michael, W. R. Schucany, and R. W. Haas. Generating random variates using transformations with multiple roots. The American Statistician,
30(2):88–90, 1976.
[14] M. E. Muller. Some continuous Monte Carlo methods for the Dirichlet
problem. Ann. Math. Statist., 27:569–589, 1956.
[15] V. Strassen. Almost sure behavior of sums of independent random variables
and martingales. In Proc. Fifth Berkeley Sympos. Math. Statist. and Probability (Berkeley, Calif., 1965/66), pages Vol. II: Contributions to Probability Theory, Part 1, pp. 315–343. Univ. California Press, Berkeley, Calif.,
1967.
18